See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/259096260

Anomaly Detection for Visual Analytics of Power Consumption Data

Article  in  Computers & Graphics · February 2014

DOI: 10.1016/j.cag.2013.10.006

CITATIONS READS

41 1,017

4 authors, including:

Halldor Janetzko Florian Stoffel

Lucerne University of Applied Sciences and Arts LKA NRW
51 PUBLICATIONS   619 CITATIONS    31 PUBLICATIONS   281 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

BigGIS View project

Visual Analytics View project

All content following this page was uploaded by Halldor Janetzko on 29 August 2018.

The user has requested enhancement of the downloaded file.

 Anomaly Detection for Visual Analytics of Power Consumption Data

Halldór Janetzko, Florian Stoffel, Sebastian Mittelstädt, Daniel A. Keim

University of Konstanz. Email: firstname.lastname@uni-konstanz.de

Abstract
Commercial buildings are significant consumers of electrical power. Also, energy expenses are an increasing cost factor. Many
companies therefore want to save money and reduce their power usage. Building administrators have to first understand the power
consumption behavior, before they can devise strategies to save energy. Secondly, sudden unexpected changes in power consump-
tion may hint at device failures of critical technical infrastructure. The goal of our research is to enable the analyst to understand
the power consumption behavior and to be aware of unexpected power consumption values. In this paper, we introduce a novel
unsupervised anomaly detection algorithm and visualize the resulting anomaly scores to guide the analyst to important time points.
Different possibilities for visualizing the power usage time series are presented, combined with a discussion of the design choices
to encode the anomaly values. Our methods are applied to real-world time series of power consumption, logged in a hierarchical
sensor network.
Keywords: importance-driven, pixel-based visualization, anomaly detection, visual analytics

1. Introduction on a time-dependent energy consumption model. We have ex-

Commercial buildings consume a significant amount of elec-
tricity. According to the Energy Information Administration’s
2010 statistics [1], the United States alone consumed an esti-
mated 1.3 trillion kW. It is about 37% of the total electricity
generated. How power is used in a commercial building has
a large effect on energy efficiency strategies. The most impor-
tant energy usage is lighting. Then heating and cooling are next
in importance [2]. Current approaches for reducing the power
consumption for example integrate motion detection sensors for
each lamp switching them on and off.
There is a growing interest in understanding how energy is
spent in the commercial buildings. Furthermore, building ad-
ministrators want to know how to reduce the failure rate and
detect anomalies. In addition, they want to know how to vi-
sualize large volumes of energy consumption data collected by
power meters (sensors) in a building to find patterns, trends,
and anomalies. In the end, our goal is to find how to automat-
ically discover the anomaly, like unusual power consumption
measurements highly differing from old observed patterns, and
to reduce the energy cost of a building. For this task, anoma-
lies are of special interest, because they can be caused either by
faulty equipment or potentially misconfigured devices consum-
ing significantly more or less energy than required for proper
operation.
In this paper, we present an analytical and visual approach
to support the building administrators in detecting anomalies
and examining energy consumption data as shown in Figure
1. Our input data consists of a tree of time series reflecting
the hierarchical nature of the power meters, e.g., one meter for
the whole building and one for each power outlet. The an-
alytical part is the automatic anomaly detection and is based
Preprint submitted to Computers & Graphics
plored two different anomaly discovery methods. In the begin-
ning, we estimate the error rate using prediction. Then, we use
clustering-based anomaly detection. Both methods have bene-
fits and drawbacks and are complementing each other.
The last step in our pipeline is the visualization being capa-
ble of effectively displaying large amounts of data and, at the
same time, allowing quick recognition of anomalous regions
in the data. We integrated the three most common time series
visualization techniques (line charts, spiral visualizations, and
Recursive Patterns) presented in Aigner et al.’s book about time
series [3]. Besides giving an appropriate overview of the data,
the visualization is also able to support the administrator in a
more detailed examination of the data, for example areas with
unusual power consumptions by interaction facilities. In ad-
dition, the visualization is capable of showing the hierarchical
nature of the data set. This is necessary, because commonly
the energy consumption of different floors or buildings is inde-
pendently monitored resulting in an inherent hierarchy in the
recorded data.
Our methods rely purely on the recorded power consump-
tion data, which we were not cleaning in any way as the data
was in very good shape. There are many external influences
to the power consumption, like the environmental conditions
or the number of people working in an office building. The
large number and high complexity of external factors prohibit
the fully automatic diagnosis of anomalies. Hence, a human
subject matter expert is needed to validate found anomalies and
investigate the interesting ones. Even though it is possible to
think of extensions for an automatic analysis of anomalies like
incorporating external factors as weather data and holidays.
It is important to note that our methods are applicable not
only to power consumption time series data sets, although they
October 4, 2013
 Time series Anomaly
detection
Figure 1: The input set of hierarchical time series is processed by anomaly detection methods. The resulting anomaly values are visualized together with the time
series values by pixel-based techniques. The visualization combines the raw time series with boosting techniques like highlighting and blurring for the anomaly
scores.
have been developed with a particular application in mind. This
is caused by the general nature of time series data and the gen-
erality of both, the analytical and the visual methods presented
in this paper. The most application-dependent part of this work
is the anomaly detection being designed for daily patterns.
2. Related Work
Reading energy consumption statistics shows that commer-
cial buildings have a high energy usage, which motivates many
research projects developed to improve power efficiency. Within
the context of our work two main categories can be distinguished:
analysis of power consumption data (detecting whether the en-
ergy consumption performs normally or abnormally over dif-
ferent locations and time) and visual analysis (visualizing simi-
larities and anomalies with appropriate interaction techniques).
2.1. Analysis of Power Consumption Data
Applying data mining techniques for power consumption
data is a known approach for identifying abnormal usage be-
havior. Agarwal et al. [4] examined 6 months of data from
the UCSD campus, including aggregate power consumption of
four buildings. Agarwal et al. focus more on the setup of power
meters and provide only simple visualization methods like line
charts. Catterson et al. [5] used an approach to monitor old
power transformers. Their goal is to proactively search for ab-
normal behavior that may indicate the transformer is about to
fail. Similarly, McArthur et al. [6] searched for anomalies to
detect problems with power generation equipment. Jakkula and
Cook [7] compared several outlier detection methods to find
which is better at identifying abnormal power consumption.
Seem [8] used outlier detection to determine if the energy con-
sumption for a day is significantly different from previous days’
energy consumption. This is a known approach for identifying
abnormal system behavior.
The work conducted at Lawrence Berkeley National Labo-
ratory [9] focuses on demand response. Mathieu et al. used a
2
Pixel-based time series &
anomaly visualization
time-of-week and piecewise-linear modeling approach to ana-
lyze commercial and industrial electric load data. To our knowl-
edge, the unsupervised anomaly detection algorithms from pre-
diction and clustering described in this paper differ from the
Mathieu et al. method in two aspects: finer granularity and
weighted by time distance (recent data weights more than old
data).
The review of several prediction methods for power data
performed by Zhao et al. in [10] investigates the effectivity
and efficiency. Neural networks and Support Vector Machines
were performing better than statistical approaches. We though
decided to use the prediction technique developed in [11] as
peak-preservation is one of the main strengths of this technique.
2.2. Visual Analysis
Visualization of building energy consumption has not yet
been a major focus of research thus far. Most of the energy
consumption visualizations have been time series line charts,
scatter plots, and maps [12, 13, 14, 15]. Recently, Many Eyes
[16] allows analysts to choose a visualization type for analyz-
ing public building electricity consumption. The Google Pow-
erMeter [17] recently provides a free energy monitoring tool for
people to view home energy usage.
In addition to these existing tools, improving visualization
techniques for time series data is ongoing research work. In
SAVE [18], Lei Shi et al. presented a sensor anomaly visualiza-
tion engine that guides the user to diagnose sensor network fail-
ures and faults using multiple coordinated views. In this paper,
we map multiple sensors’ calendar time series in a single view
to enable users to visually analyze energy usage and identify
anomalies. In SAGA Dashboard [19], Buevich et al. provided
a visual interface for interaction with the sensor network. They
require the user to use a device that tracks and visualizes home
energy usages. We extend the home energy consumption vi-
sual analysis to large commercial buildings with dozens of sen-
sors. We therefore restricted ourself to space-efficient visualiza-
tions like pixel-based Recursive Patterns. Furthermore, no pre-
defined devices and sensor types in our methods are required.
Another related work being capable of visualizing hierarchical
time series data are the TimeEdgeTrees introduced by Burch
and Weiskopf in [20]. The technique shows the time series as
one-dimensional, color-coded timelines instead of drawing the
graph edges. The hierarchy is preserved better by this approach
while the space-efficiency is worse compared to the pixel-based
approaches we use. We chose the pixel-based techniques as pe-
riodic patterns are easier perceivable. Additional discussions
on related work concerning anomalies detection and boosting
methods can be found in sections 3 and 4.
2.3. Our contribution
To leverage the prior work and to support analysts in un-
derstanding power consumption data, we combine automated
anomaly detection algorithms with interactive time series vi-
sualizations. The resulting anomaly score is used to highlight
unusual power usages in the time series visualizations. Our con-
tributions in the visual analysis process of power consumption
data are:
1. In the anomaly detection process, we
• detect power consumption anomalies based on ei-
ther a clustering-based approach or a time-weighted
prediction.
• compare the prediction-based method with a simi-
larity based anomaly computation.
2. In the time series and anomaly visualization process, we:
• map the hierarchical time series onto a Treemap.
• embed in each Treemap cell the corresponding me-
ter’s time series visualization.
• provide different time series visualization techniques
dependent on the analysis purpose.
• visualize the anomaly score by visual boosting of
the raw time series representation.
Furthermore, we have provided an advanced visual inter-
face enabling the user to visually analyze the power usage. His-
tograms for viewing the frequency and power usages of impor-
tant meters; visual queries for analyzing correlation and similar-
ity; and various options on visualization types, Treemap layout,
colormappings, and anomaly score computations enable the an-
alysts to tailor the visualization to their needs.
3. Anomalies Detection
Detecting and exploring of anomalies in time series is a
very important aspect, especially when dealing with power con-
sumption data of physical infrastructure. Saving cost and en-
ergy as well are one of the main motivations for observing and
analyzing consumption data. But when dealing with infrastruc-
ture that may be even system-critical the number of failures
must be reduced to an absolute minimum. Early signs of fail-
ure should be visible in abnormal power usage patterns. In our
3
main usage scenario abnormal behavior is defined as a differ-
ence from the expected daily pattern. Both methods described
below assume a daily power usage pattern which, of course, can
be different for each day of the week. Both techniques are not
limited to daily patterns, but can be easily adapted to the peri-
odicity of the underlying data set. The first described method
is based on a weighted prediction, where recent measurements
have a higher impact than older measurements. The latter ap-
proach is transforming the observed daily pattern in the fre-
quency domain and looking for dissimilarity in a transformed
space.
3.1. Prediction-based Anomaly Detection
The basis for prediction is an observed pattern and the as-
sumption that it is reoccurring (with slight modifications) in the
future. If this assumption does not hold true, the predicted val-
ues may be far off the measured values. Considering this fact
the other way round, observed values far distant from the ex-
pected ones tell us the model used does not explain the observed
values. There might be two reasons, the first one is that the
model quality is not good enough and the second one is that the
values are really differing from the expected and explainable be-
havior. We assume our data follows a regular underlying pattern
and therefore also assume that the model describes the usual be-
havior well. Detecting anomalies using prediction follows this
idea and is related to the statistical measure of residuals.
The prediction method used is crucial for the reliability and
expressiveness of the computed anomaly scores. As already
stated above we assumed daily patterns and include develop-
ments over time into the prediction process. We decided to
use a prediction method developed and introduced in our ear-
lier works [11]. Basically, this method predicts a value for
each minute of the day by taking all previous measurement at
the same time of the day. As an example, assume we predict
the value for a Tuesday at 11:05am. We would now average
all previous observed values of a Tuesday at 11:05am. Tak-
ing just an average would have the disadvantage of neglecting
recent developments in the time series. We therefore used a
weighted averaging scheme with higher factors for recent val-
ues and linearly decreasing influence weights for older values.
Further detailed explanations can be found in [11]. This pre-
diction method works very well for weekly patterns and will
neglect holidays or other external events. The prediction model
will adjust to seasonal changes, but alternating behaviors can-
not be modeled by this approach. Furthermore, power usage
patterns randomly distributed over a day will negatively influ-
ence the prediction quality.
After predicting for each point in a time series the expected
values based on all values occurring before this point in the
time series, we can compute the difference between predicted
and observed values. The difference is an indicator for the ab-
normality of the point in a time series but needs for higher ex-
pressiveness some kind of normalization. From the choice and
the design of the prediction method we are assuming a model
which may not being applicable to all observed time series. We
counterbalance for this fact by calculating the average fitting of
our model. More in detail, we compute the average deviation
from the predicted values for the whole time series. If a whole
0 am max max
time series is highly unpredictable, the differences between pre-

1st Monday
6 am
dicted and actual values are less meaningful compared to a case 12 pm

when a time series follows perfect daily patterns with small de- 6 pm

viations. Computation of the anomaly score is summarized by 0 am

power consumption
2nd Monday

anomaly score
6 am
the following equation: 12 pm
6 pm

0 am

3rd Monday
|predVal[time] − obsVal[time]|
anomaly[time] =
6 am
12 pm
avgt∈T ime (|predVal[t] − obsVal[t]|)
6 pm

0 am
The variable time is the point in a time series for which

4th Monday
6 am
the anomaly score is calculated. At this position the difference 12 pm

between the predicted and observed value is computed and af- 6 pm min min
terwards normalized by the average deviation from the model. a) actual values b) prediction-based
anomaly score
c) clustering-based
anomaly score

3.2. Clustering-based Anomaly Detection

The second approach for detecting anomalies in time series
data is similarity-based. We assume often-observed patterns to
be the usual behavior and rarely occurring patterns to be abnor-
mal. Following this idea, we first have to define and compute
the similarity of patterns in order to detect whether a pattern
occurs more than once. The approach described in this section
is proposed and presented by Bellala et al. in [21, 22]. The
time series is first partitioned into days and afterwards trans-
formed by a Fourier transformation into the frequency domain.
Each day of the time series is resulting in a k-dimensional vec-
tor in the frequency domain with k being a parameter of the
transformation process. The next step described by Bellala et
al. is a dimension reduction by multi-dimensional scaling into a
two-dimensional space. The density distribution in the reduced
MDS space is now interpreted as an anomaly score. Points
(time series of a single day) being in a high-density area with
many (similar) neighbors are assumed to reflect the usual be-
havior. Outliers in the 2D space can be seen as days with un-
usual values and are assigned a high anomaly score. This tech-
nique only takes the frequency domain into account and does
not integrate external effects like weather data or week of the
day.
3.3. Comparison of Anomaly Detection Methods
We previously described two methods for computing and
detecting anomalies and both come with their advantages and
drawbacks. Comparing both methods the most obvious differ-
ence is the resolution of the anomaly score. The prediction-
based method computes for each point in a time series one
anomaly score, whereas the clustering-based method returns
only one anomaly value per day. It is of course possible to ex-
tend Bellala’s technique to cope with hours or even minutes of a
day, but noise might influence the clustering approach. This be-
havior is inherited from the computation of the anomaly scores.
The clustering-based technique uses daily time series and uses
them as one data item in the clustering process. An anomaly
value is assigned to each data item based on the density dis-
tribution. Therefore, there is no possibility to assign different
anomaly scores to temporal sub-units of a day (i.e., minutes,
hours).
4
Figure 2: Comparison of the resulting anomaly scores based on the proposed
methods. The third Monday shows an unusual behavior being reflected in the
anomaly scores.
The second essential difference can be seen in the complex-
ity of the methods. Transforming the dataset in the frequency
domain and applying MDS results in a data space with axes
hard to interpret. But the frequency domain is typically less
prone to noise and induces some robustness to the observed
time series. Though the transformed data space is complex,
there exists the possibility to extract models of typical behavior
by computing cluster representatives. Furthermore, the cluster-
ing approach allows supporting several typical ’behaviors’ of
a time series. Just assume a time series alternating between a
day-work and a night-work pattern. The clustering approach
will assign both low anomaly scores as both patterns are ob-
served often, whereas the prediction-based method will assign
each day a very high anomaly score as averages are computed.
In Figure 2, we present a visual comparison between both
anomaly computation methods. The first column shows a vi-
sualization of the observed values for four consecutive Mon-
days. The exceptional behavior of the third Monday is obvi-
ous. This anomaly is reflected in all computed anomaly scores,
while the higher temporal resolution for the prediction-based
method is visible. Altogether, the clustering-based approach
is good for cases when time series switches between different
typical behaviors and the prediction-based approach is good for
cases when the behavior slightly changes over time following a
(seasonal) trend.
4. Anomalies Visualization
The anomaly scores computed in the previous section are
used to highlight important time intervals of the input time se-
ries. The visualization for the time series is influencing the de-
sign possibilities depending on the visual variable encoding the
numerical values. We implemented for comparison three well
known, state-of-the-art methods to visualize time series data:
Recursive Patterns [23, 24], Spirals [25], and the traditional
line chart. These techniques are configured to visualize only
one time series.
 We will discuss the different design alternatives and moti-
vate our design decisions in the following sections. We focus
hereby on the possibilities to encode the time series and the
anomaly values simultaneously. We describe all state-of-the art
techniques visualizing time series, namely Recursive Patterns,
spiral visualizations, and line charts.
4.1. Recursive Pattern
The Recursive Pattern is a pixel technique using coloring to
visually encode numerical values using a parameterized space-
filling layout. They are capable of displaying large amounts
of time series data in a space-efficient way. By setting proper
parameters the resulting display can be calendar-like pixel vi-
sualizations. The layout for one day is shown in Figure 2 and
the values of the series are clearly recognizable and furthermore
patterns or exceptions are easy to identify.
In order to incorporate the anomaly score in the Recursive
Patterns, we present a color boosting technique that highlights
data points by manipulating the intensity of color values ac-
cording to the anomaly score. The boosting techniques applied
in this paper are a subset of the techniques presented by Oelke
et al. in [26]. Out of the presented techniques, we used only
color highlighting as the visualization will not be overcrowded
with visual cues. Color boosting techniques bias the visual im-
pression, which might lead to misinterpretations of the visu-
alization. To keep these artifacts at a minimum, we created a
perceptual uniform colormap that only varies over hues with-
out variation in intensity. Since the change in intensity does
only minimally shift the hue, the original color tone can be re-
constructed mentally.
It is known that RGB and HSV are not perceptually uniform
and that linear interpolations within these models do not pro-
duce color scales with equal or monotonically changing light-
ness [27]. CIE LUV and CIE LAB have already been proven
useful in former visual analytics research [28, 29]. By varying
over the color opponents (a and b) but maintaining the same
lightness value L, a perceptually uniform colormap can be cre-
ated in the CIE LAB color space. However, interpolations in
CIE LAB can lead to undefined RGB signals and thus, this color
space cannot be used in the final application. Therefore, we use
the HSI color space [30] for intensity manipulation. This color
Figure 3: Different methods to display the anomaly value. Top row: the time
series values without anomaly values. Second row: the intensity of the color
is adapted to the anomaly value. Third row: color intensity representing the
anomaly score combined with adaptive Gaussian blurring.
5
space is an extension to the HSV color space that allows mono-
tonical changes in lightness.
The two proposed color encodings for the anomaly values
can be seen in Figure 3. The first row depicts the original time
series without any anomaly scores. We use different intensity
levels to encode the anomaly scores and highlight important
areas. The effect of the intensity boosting can be seen in the
second row of Figure 3. For further visual boosting we com-
bined blurring and intensity highlighting shown in the last row
of Figure 3.
We added another highlighting technique, in order to direct
the analyst to the anomalous regions of the time series. This
highlighting imitates the human perception regarding a focus
and the context area, where usually the focus area is sharp and
the context area is blurry. We used a similar approach to Kosara
et al. and Giusti et al. in [31, 32]. Since we have the anomaly
score for every element of the visualization, we are capable to
determine the important areas of the time series analytically
corresponding to the focus area of the analyst. The implementa-
tion adapts locally the blurring according to the anomaly value
of each element in the Recursive Pattern. Low anomaly val-
ues are more blurry than areas with a high anomaly score. This
adaptive blurring technique utilizes the human depth intuition
guiding the analyst to the interesting areas first in a pre-attentive
way, depicted in the bottom row of Figure 3. The blurring will
affect the visibility of pixel borders, and it influences the com-
parability between highlighted and non-highlighted areas. We
though believe that the preattentive focusing on anomalies helps
the analyst in assessing interesting points in time at a glance.
4.2. Spiral Visualization
The spiral visualization is a technique to display recurring
time series data with a fixed periodicity. Our implementation is
based on an Archimedean spiral, where the radius grows pro-
portionally to the spiral angle, which leads to a uniform expan-
sion of the spiral over time. In our implementation, each round
of the spiral is used to display one day of data. The proportional
growth of radius and spiral angle, combined with the absence
of any border between each circle makes it possible to build a
space-efficient visualization. Comparing the value of the same
time span on different days is possible, because these values are
Figure 4: Spiral visualization of time series. The left spiral shows the actual
time series data, the right spiral shows the time series data with brightness and
saturation value adapted to the anomaly score of the corresponding polygon.
on a straight line going from the center of the spiral to the out-
ermost part of the spiral. Each polygon along this line displays
the same time span of different days.
To show the anomaly score of each of the displayed time
spans, we apply the same color manipulations as described for
the Recursive Pattern above. The right spiral in Figure 4 shows
the described color saturation and brightness adjustment to high-
light the anomalous values of the time series. By comparing
the left with the right spiral the highlight of the outer ring of
the right spiral is clearly visible. There is a time range with
unusual numerical values beginning after one fourth of the day
and lasting for one quarter of a day. Besides that, some little
colorful spots are visible in the right visualization, which were
not that visible when applying only the brightness or saturation
modification technique.
4.3. Line chart
The most common visualization of time series data is un-
doubtedly the line chart. The main difference to the Recursive
Pattern or spiral-based visualization can be found in the encod-
ing of the actual time series value. In the latter two, the series
value is shown by colored polygons, which have a spatial ex-
tent. In contrast, encoding the value in a line chart is done
by the position on the y-axis. The brightness and saturation-
based techniques to add the anomaly value into the visualiza-
tion makes no sense in such a positional encoding, having only
a very small area available for the coloring. Coloring segments
of the line and applying the same techniques to enrich the line
with anomaly score information as before is not helpful as line
segments are very hard to see. To use coloring a larger line
stroke would be necessary, which would introduce high amount
Figure 5: The line chart visualization in a Tree Map with a horizontal strip lay-
out using the value weighted anomaly to determine the cell size. The effect of
the layout score is clearly visible enlarging the time series not with the high-
est anomaly score, but with the highest anomaly level regarding the time series
values.
6
of over-plotting and visual clutter. It may be fine for line chart
displayed on a large screen, but as soon several line charts are
displayed the technique does not work anymore.
To show the anomaly value simultaneously with the time
series values, we used the empty space in the background of the
line chart as shown in Figure 5. For each data point, we plot a
red stripe in the background. The anomaly value is mapped to
the opacity of the stripe in a way, that for the lowest anomaly
value it is completely transparent and therefore not visible. In
contrast, the highest anomaly score causes the stripe to have the
highest opacity resulting in a clearly visible, red stripe.
To reduce the visual clutter introduced by coloring the back-
ground, we also support a minimized view. In this view, the
anomaly stripes are only plotted above and below the line chart,
which keeps the visualization distraction-free, but still shows
the anomaly values. A comparison of both anomaly visualiza-
tion techniques for line charts can be seen in Figure 6.
Figure 6: Comparison of the anomaly visualization technique for line charts.
On the left, the whole background is used to show the anomaly scores, whereas
on the right, only a small stripe on the top and bottom of the chart background
is used to display the anomaly score, which reduces the clutter from the back-
ground coloring.
4.4. Treemap Integration
We integrated all visualizations in a Treemap display [33,
34, 35, 36] (see Figures 5 and 7). In that way, the hierarchical
nature of our time series dataset is reflected in the visualiza-
tion. Treemaps are showing the leaves of each selected branch
and the nesting depth by borders. The selection of visualized
nodes can be achieved twofold, either by interactive roll-up or
drill-down operations in the Treemap visualization or by an ad-
ditional vertical tree representation. Our design choice using
Treemaps, though they visualize only the leaves of each branch,
was implied by the application needs. The analysts are mainly
interested in finding the root-causes of anomalies first and later
on in analyzing the impacts by traversing the hierarchy to the
root node. Further details concerning the used time series can
be seen in the application section 5.
Each cell of the Treemap contains the visualizations of the
time series building one branch of the hierarchy. The border
of each of the cells is furthermore drawn in white to allow a
clear distinction in terms of the hierarchy. The caption of each
Treemap cell is used to display the numerical value used for
layout and the cell label.
The numerical value is used by the layout manager to com-
pute the final Treemap layout and directly influences the size of
a single Treemap cell. The computation of the numerical values
is critical for the expressiveness of the visualization since the
size of a cell has a large influence on the perception. The size of
a Treemap cell can be computed by different measures. Given
the interest of an analyst to quickly recognize unusual or highly
anomalous time series, the Treemap layout can be adjusted to
support these tasks by computing the layout score in different
ways. For example, the analyst can choose between the sta-
tistical variance, sum, or the arithmetic mean of the anomaly
score. To incorporate the level of the anomaly, there is also the
possibility to compute the layout based on the product of the
anomaly score and the time series value. In addition to anomaly
score-based layouts, the sum and the statistical variance of the
time series values can be used to compute the layout. Having
those choices, the visualization can be adapted to the priorities
of the analyst independently of the visualization technique. We
also added the possibility to assign the same importance value
to each node resulting in a regular layout enabling easy compar-
isons. Beside the general layout the actual width and height (the
aspect ratio) of a single cell is an important factor when using
different time series visualization techniques. For that reason,
we implemented different layout algorithms for the previously
described visualization methods.
Figure 7: Treemap visualization of 19 time series, each time series has four
weeks of data. Interesting spots or patterns in the data are highlighted and can
be therefore easily detected.
A Recursive Pattern has a rectangular shape and, therefore,
a squarified layout [37] is applied to the Treemap. This layout
algorithm results in a square-like cell, which obviously leads
to an efficient space usage of the overall display. In addition,
we framed the Treemap cells to improve the overall structure
perception of the Treemap and the hierarchical representation.
The circular shape of the spiral graphs combined with the
squarified Treemap layout leads to the best readability and space
efficiency. We hereby maximize the size of visualization and at
the same time use as much space of the Treemap cell as pos-
sible. Creating the layout for Treemaps containing line charts
comes with a fundamental difference to the Recursive Pattern
and the spirals: the width of a line chart is much larger than its
height as our observed time span is quite long. Consequently,
7
this leads to the conclusion that a squarified layout is not the
best choice. Instead, we implemented a so called strip lay-
out [38], which makes sure that the line charts are getting more
space on the horizontal axis than on the vertical (see Figure
5). Otherwise, the line charts would be very hard to interpret
and this would be an unfair comparison to the pixel-based tech-
niques. Note that the size of each Treemap cell still reflects the
numerical value used for layout.
4.5. Comparison of Anomaly Visualizations
We have presented three different state-of-the-art visualiza-
tion approaches for time series and visual extensions to show
time series and anomaly score simultaneously. All techniques
have their own advantages and disadvantages. The Recursive
Patterns presented first have the ability to visualize large amounts
of data in a very compact and space efficient way. Regardless of
the shown time range, lasting from weeks and months to years,
the Recursive Patterns are always capable of showing the data
in a readable fashion revealing patterns. The visualization is
designed in such a way that the value representation by color
enables the analyst to easily spot interesting areas or regular
patterns, nearly independent of the actual size of the visualiza-
tion. In Figure 7, patterns and outstanding time spans are visi-
ble, even in the compact Treemap representation of 19 different
time series. Having spotted regular patterns Recursive Patterns
enable also the cross-comparison in different time series, since
the relative position of one point in time is well-aligned. Using
Recursive Pattern in Treemap is more difficult to compare the
same hour of a day, for example, as the position of the same
hour varies through the visualization.
Comparing the same hour is an advantage of the spiral vi-
sualization as the periodicity was set to daily patterns. The an-
gular encoding of the time of a day enables these comparisons
as a straight line from the spiral center to the outer spiral con-
nects these data values. With such visualizations, it is easy to
Figure 8: Screenshot of our prototype showing the hierarchical and temporal
selection possibilities together with the visualization panel.
Figure 9: Overview of the power consumption data from 28 sensors during 48 weeks. Despite the huge amount of data, patterns are still clearly visible. On the
right, the same visualization with adaptive blurring highlighting unusual power consumptions can be seen.
explore the value of the time series over time. In addition, com-
paring time ranges and/or spot longer lasting trends is a simple
task, since the analyst has only to follow the continuous spiral
over time. This is an advantage compared to the non-continuous
time display of the Recursive Patterns, where layout breaks are
needed, as with any space-filling curve. Line charts are great for
detailed visual explorations of continuous data for single time
series. For the usage scenario of anomaly visualization, there
exist only a small number of application possibilities, since con-
densed visualizations are needed as limited screen space is an
issue. The low space efficiency of line charts leads to our pro-
posed solution to re-use the empty space in the background to
visually encode the anomaly value. We avoid the arising vi-
sual clutter by applying the stripe based anomaly visualization,
which keeps the anomaly information but reduces the colored
area distracting the analyst.
5. Applications
The prototype integrating all the presented analytic and vi-
sual techniques focuses especially on the detection of anoma-
lies and their temporal occurrence. With this task in mind, two
general use cases can be identified. First, general browsing and
exploration of the data is important to get an overall impres-
sion of the power usage. All different visualization techniques
presented above can be applied to gain from their individual
strengths. The second task is the examination of a specific is-
sue, like unusually high or low power consumption. Our system
can provide the analytical and visual insights necessary to find
the source of the unusual energy consumption. All visualiza-
tions are integrated in the same analytical framework, but use
different methods of displaying the power consumption and the
anomaly values.
5.1. Analytical Framework
Our prototype consists basically out of three parts reflecting
the different dimensions in the data set and can be seen in Fig-
8
ure 8. The left panel allows the navigation through the hierar-
chy of the sensor graph by selecting the nodes being visualized.
The visualization panel in the center consists of the Treemap
visualization together with a colormap legend. The panel at the
bottom of the window allows to navigate in time and select the
time range that should be visualized. This timeline visualiza-
tion shows the total amount of power usage over time in order
to give the analysts additional hints.
We implemented beside animation also interaction techniques
like dragging the selected time range (blue rectangle in the time-
line visualization) left and right causing immediate updates to
the visualization. The visualization allows basically three inter-
action possibilities. The first is a tooltip allowing to inspect the
underlying data values invoked by mouse hovering. We further-
more directly support drill-down and roll-up operations in the
Treemap visualization, allowing the analyst to keep his focus on
the visualization during traversing the sensor graph. Finally, the
analyst is able to select a region in the visualization and query
the system for similar time series sharing the selected behav-
ior by means of distance or correlation calculations. Switch-
ing the visualization technique, colormap, value normalization,
anomaly calculation, or the weights for the Treemap layout is
possible by selecting the respective option.
5.2. Visual Inspection of Anomalies
In this use case, the building administrator gets the infor-
mation, that in February 2012 the overall power consumption
and energy costs of a building was higher than expected. The
investigation starts by getting an overview and some contex-
tual information about the general energy consumption of the
building. Undoubtedly, the most suitable visualization for this
task is the Recursive Pattern Visualization, which can be seen
in Figure 9. The blurring approach at the right side highlights
the anomalies further compared to the left figure, where we vi-
sualized anomalies only by color intensity. The resulting visu-
alization points directly to one time series, which can be seen
Figure 10: Sensor readings of sensor AE3 from 6th February to 4th of March 2012. On the left, the power consumption is visualized. On the right, the intensity of
the colors reflects the anomaly score. Due to the high intensity, an area in the fifth column of the third row stands out.

in Figure 10 on the right. Both, the left and the right visualiza-
tion show the power consumption data beginning on 6 February
2012. Each of the bigger rectangles contains the data from one
day, starting with Monday on the left. In total, there are four
weeks of data visible, starting on 6th February and ending on
4th of March.
In the visualization, there are some single, outstanding spots.
Those look relatively random and last only one pixel, which
stands for a time span of five minutes. Although the color is
quite intense and reddish, they are far too few and do not last
long enough to have a large influence on the power consump-
tion. Besides these spots, an area in the fifth column of the third
row stands out. The intensity seems to increase from pixel to
pixel over a long time. Having in mind, that one small black-
framed rectangle of the Recursive Pattern stands for one hour,
the anomaly score seems to increase over ten hours, until sud-
denly the anomaly score drops again. Due to the long dura- Figure 11: The time series query result window. On the top left, the query time
tion of the anomaly and the intense red color, the actual energy series is displayed, on the right the top-n query results are shown. The query
range is highlighted.
consumption in this time frame is very high. This makes this
anomaly a candidate for the cause of the higher energy costs in
February. this additional knowledge, the building administrator can con-
The building administrator found an anomaly in the given clude that the anomaly affected not only one, but at least four
time frame with the Recursive Pattern visualization. To identify parts of the building, where the sensors have been installed.
potentially correlated time series, our prototype implements a The quality of the conclusions drawn from the visualiza-
top-n time series similarity search. The query can be created by tions and analytical methods depends heavily on the sensor de-
clicking on a part of the visualization and selecting the query ployment. If each of the sensors monitors a single machine
area with the mouse. Afterwards, the desired similarity mea- or office, the building administrator has a concrete subject of
sure can be selected. The system supports the standard Eu- further examination. When they are deployed in a more gen-
clidean distance and positive, negative, and unsigned Pearson eral way, for example per building floor or even per building,
Correlation for different analysis tasks. In this case, selecting the shown analysis allows narrowing down the investigation of
the positive Pearson Correlation and the Euclidean Distance is power consumption to the affected units.
appropriate. The result of the query can be seen in Figure 11.
The query results show three very similar series: AE4, AE5,
6. Evaluation
and AE6. All three sensors are part of the same subtree of the
sensor hierarchy. This means they are located in the same build- We showed the applicability of our proposed technique in
ing as sensor AE3, which logged the time series identified as the previous application section, but it is very important that
anomalous by the Recursive Pattern visualization before. With real expert users rate our approach effective and helpful. We
9
therefore presented our approach to the target user group in a
big company. We had contact to two analysts and interviewed
them first about their state of the art technology. The com-
pany develops sensor networks measuring the power consump-
tion for large buildings and is experienced with power manage-
ment. The current state of the art technology they are using is a
line chart based visualization. They are able to select arbitrary
time frames and inspect the temporal power distribution. Fur-
ther analysis steps are yet impossible to perform. In later meet-
ings we explained our approach to the experts and afterwards
let them interact with our system and investigate the time series
data. We asked them to describe their typical way of analyzing
data and furthermore to comment on our proposed technique by
thinking aloud using our prototype. We got very valuable and
interesting feedback from the experts regarding the benefits and
room for improvement.
First of all, they validated the temporal patterns shown in
the pixel-oriented visualization techniques with their knowl-
edge of typical power consumption patterns. Their proof-of-
concept was that the daily periodic patterns were visible at a
glance, at the same time reflecting their expectations for the
time series. After they found the patterns like low power con-
sumption at nighttime and weekends they started to look for
anomalies using our visual boosting techniques. At first, obvi-
ous patterns like holidays or the Christmas vacation have been
found. Afterwards, less obvious patterns have been investi-
gated. During their analysis, we asked the experts to comment
on our techniques and give feedback related to visualization and
analysis methods.
The first point they commented on was the helpfulness of
the overview visualization in the form of the Recursive Pat-
terns. Compared to line chart based visualizations they are very
familiar with, the calendar-like representation of the power con-
sumption was highly appreciated. Furthermore, the possibility
to interactively change the visualization type helped them a lot
to get familiar with the pixel-oriented techniques. The color-
ing of pixels was intuitive to them and they could interpret the
visualization easily.
From an analysis point of view, a very interesting point was
their comment on our prediction-based anomaly computation.
They agreed with our definition of anomaly: ”The anomalous
day is likely to deviate from the daily pattern in some way.“ As
shown above, our anomaly method is very fine grained, but to
the experts a single time spot with a high anomaly score is not
important. They were more interested in longer periods of un-
usual behavior, starting at approximately one hour duration. On
the other hand the related anomaly computation method based
on days was too coarse-grained for them for this kind of anal-
yses. An aggregation of the anomaly values might help to let
the analyst focus on the severe anomalies. The visualization
of the anomaly scores together with the time series was men-
tioned very positive, especially with respect to the Recursive
Pattern. The overview calendar-like visualization with inten-
sity highlighting and adaptive blurring let them focus on the
interesting spots. They had the impression that their attention
was guided to the anomalies, while the unimportant, common
daily patterns were pushed in the background. As soon as they
10
found some unexpected anomalies they applied further analysis
techniques.
The experts very much appreciated the possibility to select
a region in the time series and query for other similar time se-
ries. When they selected a leaf in the hierarchy of time series
they would look for the impacts of the anomaly on the parent
nodes. The other way around, querying for anomalies on higher
levels would show the root-causes for the unusual power con-
sumption.
A possibility for improvement mentioned by them is the in-
tegration of external events into the application. Sometimes
managers know in advance of extraordinary events that will
cause unusual power consumptions. It should be possible to
include this information whenever available and to reflect the
additional events in the visualization. Overall they found the in-
tegration of different time series visualization techniques com-
bined with an anomaly representation very helpful and wanted
to integrate our techniques in their management tools.
7. Conclusion
Analyzing and interpreting unusual patterns in time series
data is a very important task. In this paper, we applied novel
analysis and proven visualization techniques to a system, which
supports analysts finding those patterns in a visual way. We
supported the analysis process by computing anomaly scores
of the given time series data with an anomaly detection algo-
rithm which produces very fine grained results. This also allows
the creation of detailed visualizations resulting in a fine grained
pixel-based date representation. Furthermore, the algorithm is
very efficient in terms of required computing power, because it
does not require expensive transformations nor does it rely on
elaborate analysis of the time series data.
Having the anomaly scores, different visualizations can be
used to get deep insight into the time series and the anomaly
scores, depending on the task to fulfill. Recursive Patterns gen-
erate overviews of large time spans and large amounts of data.
Spiral views provide the possibility to quickly detect and ana-
lyze periodic patterns. If the actual data values are of interest,
the classical line charts are also available for further investiga-
tions of the data set.
The double encoding of time series values and anomaly
scores is solved in different ways. The novel adaptive blurring,
which generates a focus and a context area by blurring the vi-
sualization according to the anomaly scores, guides the analyst
directly to interesting spots of the visualization. This makes
the technique a particular advantage in overview visualizations,
where irrelevant areas of the time series are losing their level of
detail by a strong blur, whereas interesting, high anomalous ar-
eas are clearly visible and attract the focus of the human eye. To
support the display of multiple visualizations, the well-known
Treemap approach is extended by layouts based on space effi-
ciency and specific visual properties of the visualization. Since
the anomaly scores determining the layout can be selected de-
pending on the analysis task, the resulting Treemap layout also
supports further analyses.
 The use case of power consumption data shows the appli-
cability of the methods shown in this paper. The general nature
of the analysis and visualization methods makes it possible to
apply these techniques to time series not only from the applica-
tion domain of power consumption data. In the future, we want
to integrate external knowledge like known events influencing
the time series like weather information. It would be also in-
teresting to automatically determine the visualization method,
colormap, and possible enhancements like the adaptive blurring
based on the displayed data.
Acknowledgment
We are thankful to Ming Hao of Hewlett-Packard Labs for
fruitful collaboration on the analysis of power consumption data.
This work has been partly funded by the German Research So-
ciety (DFG) under the grant GK-1042, Explorative Analysis
and Visualization of Large Information Spaces, Konstanz.
References
[1] United States Energy Information Administration . Annual energy review.
2010.
[2] US Department of Energy . Energy efficiency trends in residen-
tial and commercial buildings. 2008. http://apps1.eere.
energy.gov/buildings/publications/pdfs/corporate/bt_
stateindustry.pdf.
[3] Aigner W, Miksch S, Schumann H, Tominski C. Time & Time-Oriented
Data. Human-Computer Interaction Series; Springer London; 2011.
ISBN 978-0-85729-078-6.
[4] Agarwal Y, Weng T, Gupta RK. The energy dashboard: improving the
visibility of energy consumption at a campus-wide scale. In: Proceedings
of the First ACM Workshop on Embedded Sensing Systems for Energy-
Efficiency in Buildings. ACM; 2009, p. 55–60.
[5] Catterson VM, McArthur SD, Moss G. Online conditional anomaly de-
tection in multivariate data for transformer monitoring. Power Delivery,
IEEE Transactions on 2010;25(4):2556–64.
[6] McArthur SD, Booth CD, McDonald J, McFadyen IT. An agent-based
anomaly detection architecture for condition monitoring. Power Systems,
IEEE Transactions on 2005;20(4):1675–82.
[7] Jakkula V, Cook D. Outlier detection in smart environment structured
power datasets. In: Intelligent Environments (IE), 2010 Sixth Interna-
tional Conference on. IEEE; 2010, p. 29–33.
[8] Seem JE. Using intelligent data analysis to detect abnormal energy con-
sumption in buildings. Energy and buildings 2007;39(1):52–8.
[9] Mathieu JL, Price PN, Kiliccote S, Piette MA. Quantifying changes in
building electricity use, with application to demand response. Smart Grid,
IEEE Transactions on 2011;2(3):507–18.
[10] Zhao Hx, Magoulès F. A review on the prediction of building energy con-
sumption. Renewable and Sustainable Energy Reviews 2012;16(6):3586–
92.
[11] Hao MC, Janetzko H, Mittelstdt S, Hill W, Dayal U, Keim DA, et al. A
visual analytics approach for peak-preserving prediction of large seasonal
time series. Computer Graphics Forum 2011;30(3):691–700.
[12] http://en.wikipedia.org/wiki/Energy_in_the_United_
States. viewed 6/17/13.
[13] IBM TRIRIGA Energy Optimization: Integrated software solu-
tion for improving buildings management and facilities opera-
tion. viewed 6/17/13. http://pic.dhe.ibm.com/infocenter/
tivihelp/v57r1/topic/com.ibm.iteo.doc/infocenter.pdf.
[14] UCEI (University of California Energy Institute), Berke-
ley, California, CA . New york city building energy map.
2007. http://www.visualizing.org/visualizations/
new-york-city-building-energy-map.
[15] Granderson J, Piette MA, Ghatikar G, Price P. Building energy informa-
tion systems: State of the technology and user case studies. Handbook of
Web Based Energy Information and Control Systems 2009;.
11
View publication stats
[16] IBM Research and the IBM Cognos software group . Many eyes: Public
building energy consumptions. 2013.
[17] Google PowerMeter. viewed 6/17/13. http://www.google.com/
powermeter/about/.
[18] Shi L, Liao Q, He Y, Li R, Striegel A, Su Z. SAVE: Sensor anomaly visu-
alization engine. In: Visual Analytics Science and Technology (VAST),
2011 IEEE Conference on. IEEE; 2011, p. 201–10.
[19] Buevich M, Rowe A, Rajkumar R. SAGA: Tracking and Visualization
of Building Energy. In: Embedded and Real-Time Computing Systems
and Applications (RTCSA), 2011 IEEE 17th International Conference on;
vol. 2. IEEE; 2011, p. 31–6.
[20] Burch M, Weiskopf D. Visualizing dynamic quantitative data in hierar-
chies. In: Proceedings of International Conference on Information Visu-
alization Theory and Applications. 2011, p. 177–86.
[21] Bellala G, Marwah M, Arlitt M, Lyon G, Bash C. Following the electrons:
methods for power management in commercial buildings. In: Proceed-
ings of the 18th ACM SIGKDD international conference on Knowledge
discovery and data mining. ACM; 2012, p. 994–1002.
[22] Bellala G, Marwah M, Arlitt M, Lyon G, Bash CE. Towards an un-
derstanding of campus-scale power consumption. In: Proceedings of
the Third ACM Workshop on Embedded Sensing Systems for Energy-
Efficiency in Buildings. ACM; 2011, p. 73–8.
[23] Keim DA, Ankerst M, Kriegel HP. Recursive pattern: A technique for
visualizing very large amounts of data. In: Proceedings of the 6th confer-
ence on Visualization’95. IEEE Computer Society; 1995, p. 279–86.
[24] Lammarsch T, Aigner W, Bertone A, Gartner J, Mayr E, Miksch S, et al.
Hierarchical temporal patterns and interactive aggregated views for pixel-
based visualizations. In: Information Visualisation, 2009 13th Interna-
tional Conference. IEEE; 2009, p. 44–50.
[25] Weber M, Alexa M, Müller W. Visualizing time-series on spirals. In:
proceedings of the IEEE Symposium on Information Visualization. 2001,
p. 7–13.
[26] Oelke D, Janetzko H, Simon S, Neuhaus K, Keim DA. Visual boosting in
pixel-based visualizations. Computer Graphics Forum 2011;30(3):871–
80.
[27] Keim DA. Designing pixel-oriented visualization techniques: Theory and
applications. Visualization and Computer Graphics, IEEE Transactions
on 2000;6(1):59–78.
[28] Healey CG. Choosing effective colours for data visualization. In: Visual-
ization’96. Proceedings. IEEE; 1996, p. 263–70.
[29] Wang L, Kaufman A. Importance driven automatic color design for direct
volume rendering. Computer Graphics Forum 2012;31(3pt4):1305–14.
[30] Keim DA, Kriegel HP. Issues in visualizing large databases. In: Proc.
Conf. on Visual Database Systems (VDB’95), Lausanne, Schweiz. 1995,
p. 203–14.
[31] Kosara R, Miksch S, Hauser H. Semantic depth of field. In: Proceed-
ings of: IEEE Symposium on Information Visualization 2001 (INFOVIS
2001). IEEE Computer Society Press; 2001, p. 97–104.
[32] Giusti A, Taddei P, Corani G, Gambardella L, Magli C, Gianaroli L. Ar-
tificial defocus for displaying markers in microscopy z-stacks. Visualiza-
tion and Computer Graphics, IEEE Transactions on 2011;17(12):1757–
64.
[33] Johnson B, Shneiderman B. Tree-maps: A space-filling approach to the
visualization of hierarchical information structures. In: Visualization,
1991. Visualization’91, Proceedings., IEEE Conference on. IEEE; 1991,
p. 284–91.
[34] Shneiderman B. Tree visualization with tree-maps: 2-d space-filling ap-
proach. ACM Transactions on graphics (TOG) 1992;11(1):92–9.
[35] Schreck T, Keim DA, Mansmann F. Regular TreeMap Layouts for Visual
Analysis of Hierarchical Data. In: Proceedings of the Spring Conference
on Computer Graphics (SCCG’2006). Casta Papiernicka, Slovak Repub-
lic: ACM Siggraph; 2006, p. 184–91.
[36] Hao MC, Dayal U, Keim DA, Schreck T. Importance-driven visualization
layouts for large time series data. In: Information Visualization, 2005.
INFOVIS 2005. IEEE Symposium on. IEEE; 2005, p. 203–10.
[37] Bruls M, Huizing K, van Wijk J. Squarified treemaps. In: Proceedings
of the Joint Eurographics and IEEE TCVG Symposium on Visualization
(VisSym 00). Eurographics Association; 2000, p. 33–42.
[38] Baker MJ, Eick SG. Space-filling software visualization. Journal of Vi-
sual Languages and Computing 1995;6(2):119–33.